ST-DISTAL: Dual-Branch Graph Convolution with Distributional Alignment for Cell-Type Deconvolution

Abstract

Spatial transcriptomics enables high-throughput gene expression profiling while preserving spatial information, offering valuable insights into tissue architecture and cellular organization. Methods that achieve single-cell or subcellular spatial resolution typically rely on predefined gene panels, which limits genome-wide discovery. In contrast, sequencing-based spatial transcriptomics platforms provide broad transcriptome coverage but measure expression at lower spatial resolution, capturing mixtures of multiple cell types within each spatial location and thereby requiring accurate cell-type deconvolution. Many existing deconvolution methods do not jointly model molecular expression similarity and spatial context, fail to align predicted cell-type compositions with biological priors, and may produce spatially inconsistent results that do not reflect the smooth and structured organization of real tissues. We present ST-DISTAL, a dual-branch graph convolutional framework for cell-type deconvolution in spatial transcriptomics data. ST-DISTAL integrates conventional spatial graph convolution with spectral Chebyshev filtering through an adaptive attention mechanism, enabling the model to capture both local spatial dependencies and global structural patterns. The framework is trained using a composite objective that jointly enforces accurate cell-type proportion estimation, global distributional alignment with reference profiles, and spatial smoothness. Evaluations on three simulated benchmark datasets demonstrate that ST-DISTAL consistently outperforms state-of-the-art methods in both prediction accuracy and distributional alignment. Further validation on a real human embryonic heart dataset and a colorectal cancer dataset shows biologically plausible and spatially coherent cell-type organization.